Research On Test And Evaluation Method Of Intelligent Driving Systems Based On Naturalistic Driving Data

In the past decades,the technology of intelligent driving has made remarkable progress.Experts,scholars and relevant institutions from domestic and overseas have developed many intelligent driving systems to give modern vehicles multiple functions,making them more intelligent and safer.The development and verification of intelligent driving system are always complementary to each other.Before they’re formally applied,they need to go through numerous tests and give quantitative evaluation of their performances.At present,the research on the test and evaluation of intelligent driving system has the limitations of single test scenario for multi-function tests,besides,most of the test methods are adopted from the international standard test methods,which are not completely consistent with the actual road conditions in China.This thesis constructs typical virtual test scenarios through data-driven clustering analysis,which the datas are from naturalistic driving datas collected on the Chinese roads,and,the data inductive learning methods of Markov Monte Carlo simulation are used to establish a stochastic prediction model of the preceding vehicle in the test scenario.The intelligent driving system evaluation model is constructed from three aspects:safety,comfort and economy.Finally,based on the combination of analytic hierarchy process and fuzzy synthetic evaluation,the intelligent driving system is comprehensively evaluated.The main research points of this thesis are presented as following:(1)Test scenarios of intelligent driving system based on natural driving data are constructed.The naturalistic driving datas of the automobile under the road traffic environment in China are collected,and the dangerous working condition data fragments are selected from the naturalistic driving datas.Based on this,the scenario feature elements of the comprehensive test of the intelligent driving system are extracted,and three typical models are obtained by cluster analysis.The virtual test scenarios,which build the scenario foundation for the virtual tests of the intelligent driving systems,solves the problem that there are no comprehensive test scenarios of the localized intelligent driving system in China.(2)A stochastic motion prediction model based on markov monte carlo simulation is established in virtual test scenario.The markov chain theory was used to characterize the stochastic motion of the vehicle driven by the human driver in front of the subj ect vehicle.The subject vehicle data in each scenario obtained by clustering was used as the historical working condition data of the preceding vehicle.The markov chain transfer probability was concluded and learned,and the state at the future time was predicted by monte carlo simulation.Based on this,the stochastic motion models of the preceding vehicle in the test scenarios are obtained.The validity of the models is verified by comparing the original working condition data.The problems that the stochastic motion of the vehicle driven by human drivers could not be accurately represented in the virtual environment test due to the low accuracy of the acquisition equipment and inaccurate data of the preceding vehicle are effectively solved.(3)A fuzzy synthetic evaluation method of intelligent driving system based on analytic hierarchy process is proposed.Based on the PreScan simulation platform,the multi-case virtual tests are carried out.The multi-level evaluation model is built with the safety,comfort and economy of the vehicle as the index.The virtual test datas are used as the reference for setting the evaluation model membership degree and fuzzy rules,and the scores of safety,comfort and economy are calculated to verify the feasibility of the proposed fuzzy synthetic evaluation method for intelligent driving system.